Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter
Episode Details
The podcast features a dynamic conversation among Jason Calacanis, Chamath Palihapitiya, David Sacks, Gavin Baker, and Travis Kalanick. They begin by analyzing the recent political sweep in New York City by the Democratic Socialists of America (DSA), driven by figures like Zoran Mamdani. The hosts debate the societal implications, contrasting the DSA's platform with the potential of AI (Artificial Intelligence) as an economic equalizer. The discussion shifts to the global stage, specifically the technology race with China. The panel observes that Chinese firms are catching up rapidly in Open source AI. For instance, Z.AI recently released the highly capable GLM 5.2 model, which heavily relied on Distillation (AI) techniques and hardware from Huawei, similar to how DeepSeek operates. This rapid progress challenges US incumbents like OpenAI and anthropic. The latter, led by CEO Dario Amodei, recently saw its Fable 5 model delayed due to regulatory caution surrounding Cybersecurity in the AI era. The panel predicts a shift toward Composable Models, which combine open-source solutions with proprietary models, thereby expanding the entire AI Infrastructure market currently dominated by Nvidia and its CEO Jensen Huang. The panel then dissects the hardware bottleneck. Micron recently smashed earnings expectations thanks to massive demand for High-bandwidth memory (HBM). Alongside SK Hynix and Samsung, Micron is struggling to meet demand, while Chinese manufacturer CXMT tries to capture the lower-end market. This memory crunch is drastically raising costs for consumer hardware companies like Apple. Furthermore, the industry is grappling with severe Energy constraints in AI that limit the build-out of terrestrial Data Centers. To circumvent this, Elon Musk is exploring alternative solutions. With SpaceX and its reusable Starship rocket, he aims to drastically lower launch costs, making the deployment of data centers in space economically viable. Concurrently, Tesla recently trademarked Megapod, a modular data center concept intended for rapid terrestrial deployment. Travis Kalanick, sharing insights from his new venture Adams, joins the hosts in exploring the potential of Distributed inference. By utilizing decentralized networks like Bit Tensor Tao, unused consumer compute can be pooled. This requires disaggregating the inference process into Prefill (AI) and Decode (AI) workloads. Specialized hardware companies like Groq and Cerebras (which rely on TSMC for silicon wafers) are perfectly positioned to optimize this decode phase. Finally, the hosts analyze the IPO Market and the capital dynamics for hyper-scalers like CoreWeave and broader Cloud Computing ecosystems. Following Cerebras's post-IPO price drop, they discuss the merits of a Dutch auction to properly price highly anticipated tech offerings without breaking deal prices.
The episode explores the intersection of political shifts, the global AI arms race, and the physical infrastructure bottlenecks defining the current tech cycle. Key themes include the rise of the Democratic Socialists of America (DSA), the rapid advancement of Chinese open-source AI models, and the critical role of high-bandwidth memory (HBM) and energy availability in scaling AI data centers.
Portfolio lens: This set of ideas represents an AI infrastructure and hardware bottleneck thesis, focusing on the physical and logistical constraints of the AI revolution.
Generated with gemini-3.1-flash-lite on 6/27/2026, 5:21:59 AM. For research only. Not financial advice.HBM and Memory Infrastructure
High-bandwidth memory (HBM) is the most critical and constrained bottleneck in the AI hardware stack, creating a durable pricing power advantage for the few manufacturers capable of producing it.
Micron's earnings performance and the panel's consensus that memory capacity is foundational to AI performance suggest that HBM demand will remain price-insensitive while supply remains limited to three global players.
- Micron's revenue grew 4x year-over-year with sold-out 2026 supply.
- DRAM is projected to be 30-40% of hyperscaler capex next year.
- Only three companies globally can manufacture HBM, making it a highly specialized, non-commodity component.
- Continued supply-demand imbalance in HBM.
- New capacity coming online from major players.
- Potential market entry of Chinese manufacturer CXMT in lower-end segments.
- Potential for sudden supply gluts if capacity ramps faster than expected.
- Regulatory or geopolitical hurdles in building new fabrication plants.
- Demand destruction in consumer electronics due to component price inflation.
- Monitor HBM capacity expansion timelines and capital expenditure reports from major memory manufacturers.
- Track price trends in consumer electronics as a proxy for DRAM availability.
Composable AI Model Architectures
The future of enterprise AI is not a single frontier model, but a 'council of LLMs' where proprietary and open-source models are composed to optimize for cost, accuracy, and specific task requirements.
The panel argues that open-source models are catching up to frontier capabilities, leading to a shift where enterprises will use cheaper open-source models for routine tasks and reserve expensive frontier models for complex reasoning.
- Chinese model GLM 5.2 demonstrates frontier-class performance at a fraction of the cost.
- Enterprises are increasingly adopting 'router' architectures to direct queries to the most efficient model.
- Increased adoption of open-source models by enterprises.
- Continued performance improvements in open-weight models.
- Development of specialized 'router' software to manage model orchestration.
- Frontier labs may maintain a significant lead in reasoning capabilities.
- Regulatory or security concerns regarding the use of open-source models in sensitive enterprise environments.
- Analyze the adoption rate of open-source vs. proprietary models in enterprise software stacks.
- Evaluate the performance of model routing software.
Orbital and Modular Compute Infrastructure
As terrestrial energy and regulatory constraints make data center build-outs increasingly difficult, modular and space-based compute solutions will become economically viable alternatives.
The panel highlights that the cost of terrestrial data center build-outs is becoming inflationary due to energy, labor, and regulatory hurdles, creating a potential opening for modular pods and orbital compute.
- Tesla's 'Megapod' trademark suggests a move toward modular, rapid-deployment data centers.
- SpaceX's Starship reusability aims to lower launch costs, potentially making orbital compute economically competitive with terrestrial build-outs.
- Data center projects are increasingly facing regulatory and energy-related contestation.
- Successful deployment of modular data center units.
- Advancements in Starship reusability and launch cost reduction.
- Increased regulatory difficulty for terrestrial data center construction.
- Technical challenges in cooling and maintaining hardware in space.
- Latency issues inherent in orbital compute.
- High initial capital requirements for space-based infrastructure.
- Monitor filings and public announcements regarding modular data center deployments.
- Track Starship launch cost metrics and orbital compute feasibility studies.
Watchlist
- Micron (MU) HBM capacity and pricing
- Huawei Ascend chip production metrics
- Starship launch frequency and cost per kg
- Data center energy consumption and regulatory approval rates
- Cerebras (CBRS) and CoreWeave infrastructure build-out progress
Open Questions
- How quickly can Chinese firms scale production of indigenous AI chips like the Huawei Ascend 910b?
- Will the 'Megapod' concept be used primarily for internal Tesla/SpaceX compute or as a commercial product?
- Can distributed inference networks overcome the latency and security challenges required for enterprise-grade SLAs?
- What is the true cost-benefit analysis of orbital compute compared to terrestrial data centers once launch costs are fully amortized?
Key Topics & People
The market for Initial Public Offerings which is showing signs of reopening.
Emerging technology field experiencing a massive wave of hype and high valuations despite the broader market downturn.
.jpg?width=300){kind=avatar}
.jpg?width=300){kind=avatar}
Host on the All-In Podcast who was absent during this episode.
Host on the All-In Podcast, software investor, and venture capitalist.
.jpg?width=300){kind=avatar}
Host and moderator of the All-In Podcast.
A political organization backing progressive economic policies like subsidized city grocery stores.
CEO of Anthropic, criticized for allegedly seeking an FDA-like regulatory moat for AI.
Remote server infrastructure used to offload heavy computational workloads from robots.
The foundational layers of technology, such as data centers and silicon chips, for AI models.
Using outputs of a more advanced AI model to train other models.
Freely available and modifiable artificial intelligence models.
A Democratic socialist politician cited as a symbol of the DSA's rising electoral success.
A distributed crypto project providing decentralized AI compute capacity
.jpg?width=300){kind=avatar}
CEO of Nvidia, known for guiding the company's AI hardware dominance
Large scale facilities housing computer systems, currently experiencing massive demand
A framework where enterprises use a mix of frontier models and their own open-source models.
A pricing mechanism for IPOs suggested as an alternative to traditional underwriting.
Utilizing a decentralized network of hardware to run inference for AI models.
The part of AI inference focused on sequentially generating the next tokens.
The part of AI inference focused on understanding the input prompt and its context.
The concept of hosting AI data centers in orbit to circumvent terrestrial energy and zoning constraints.
The severe bottleneck in power availability necessary to run large-scale AI data centers.
Specialized memory chips essential for AI GPUs, facing severe global supply shortages.
The intersection of AI capabilities and cybersecurity, raising concerns about automated vulnerabilities.
.jpg?width=300){kind=avatar}
Entrepreneur, founder of Adams, and podcast guest discussing software, AI, and politics.
Investor and podcast guest discussing space, IPOs, and technology markets.